Notes:

---

1. Unpaired image-to-imgage translation
2. Cycle-consistency loss
3. Good result
4. Only works on texture and color translation, does not perform well on deformation.

Result:

---

Summary:

---

1. Network structure:

   

2. Objective

   • GAN loss: Traditional GAN loss

     

   • Cycle loss: Kind of like reconstruction loss, but images is generated by two generators

     

Links:

---

arxiv: https://arxiv.org/abs/1703.10593
github: https://github.com/junyanz/CycleGAN